1668766 - mach try auto didn't schedule hazard builds

Status: NEW

(Developer Infrastructure :: Try, defect, P3)

Description

[Simon Giesecke [:sg] [he/him]]

The mach try auto push https://treeherder.mozilla.org/#/jobs?repo=try&revision=7a7f8f60d33b5382e77a4c8b44f67b61ec85fd08 didn't file any hazard build AFAIU, and the subsequent landing was backed out due to a bustage of that build: https://bugzilla.mozilla.org/show_bug.cgi?id=1665347#c9

Comment 1

[Andrew Halberstadt [:ahal]]

Thanks for the report, Marco can the ML schedule hazard builds? Or is this just a case of not enough training?

Comment 2

[Andrew Halberstadt [:ahal]]

Clearing stale needinfo

Comment 3

[Marco Castelluccio [:marco]]

Sorry I had missed this... I will check.

Comment 4

[Marco Castelluccio [:marco]]

I've checked, we are indeed ignoring them. I can add them if we want to.
What are they doing exactly compared to normal builds?

Comment 5

[Simon Giesecke [:sg] [he/him]]

They are performing a GC rooting hazard analysis. sfink can provide a more substantial answer, if required.

Comment 6

[Steve Fink [:sfink] [:s:]]

Yes, it's a static analysis that requires a full browser build (with GCC). https://wiki.mozilla.org/Javascript:Hazard_Builds describes them. I'm actually not sure when it makes the most sense to trigger them. Certainly any code change within js/** could cause them to fail, but also any code change anywhere else can cause them to fail in any code that calls the changed code. (Such is the story with whole-program analysis.) But the vast majority of changes won't be relevant. I guess the failures are pretty heavily clustered around changes to certain files, so perhaps ML would do well here?

Comment 7

[Andrew Halberstadt [:ahal]]

Clearing stale needinfos (please re-request if still needed)